Optimizing parallel programs with dynamic data structures

Distributed memory parallel architectures support a memory model where some memory accesses are local, and thus inexpensive, while other memory accesses are remote, and potentially quite expensive. In order to achieve efficiency on such architectures, we need to reduce remote accesses. This is particularly challenging for applications that use dynamic data structures. In this thesis, I present two compiler techniques to reduce the overhead of remote memory accesses for dynamic data structure based applications: locality techniques and communication optimizations. Locality techniques include a static locality analysis, which statically estimates when an indirect reference via a pointer can be safely assumed to be a local access, and dynamic locality checks, which consists of runtime tests to identify local accesses. Communication techniques include: (1) code movement to issue remote reads earlier and writes later; (2) code transformations to replace repeated/redundant remote accesses with one access; and (3) transformations to block or pipeline a group of remote requests together. Both locality and communication techniques have been implemented and incorporated into our EARTH-McCAT compiler framework, and a series of experiments have been conducted to evaluate these techniques. The experimental results show that we are able to achieve up to 26% performance improvement with each technique alone, and up to 29% performance improvement when both techniques are applied together.